Answer generation device, answer generation program, answer generation method, and AI call center system

Abstract

We provide user-friendly answer generation devices, answer generation programs, answer generation methods, and AI call center systems. [Solution] An AI operator that generates response text and voice in response to utterances supplied from a user terminal and enables conversation with the user, wherein the memory unit 55 records multiple tasks to be performed during the conversation, detailed subtasks for each task, the execution order, and preset responses corresponding to each subtask. The response generation unit analyzes the meaning of the utterance by referring to the conversation history, searches for a preset response corresponding to the currently executed subtask, and if there is no appropriate response, searches for a preset response within the same task or in another task and generates the response text.

Description

Technical Field

[0001] The present invention is suitably used for a call center that uses, for example, both an AI operator and a manned call center.

Background Art

[0002] In recent years, in order to improve the efficiency of telephone-based sales, the utilization of automatically acquired CTI (Computer Telephony Integration) information such as call duration and number of calls has been progressing, and AI call center systems configured to record the information in association with inquiry information are known (see, for example, Patent Document 1).

[0003] As shown in Figure 1, in the AI call center system P1, when a call is made via a user terminal P4 owned by a user, voice information is supplied via a CTI device P2 to one selected call handling terminal P3B among a plurality of call handling terminals P3B belonging to a call center P3.

[0004] The call handling terminal P3 is operated by a human operator. The call handling terminal P3B can access a CRM (Customer Relationship Management) device P7 that manages customer's personal information and usage history, and a status management device P8 that manages service reservations and product inventory status, and executes a call with the user while checking the customer's personal information, usage history, and inventory status.

Prior Art Documents

Patent Documents

[0005]

Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

[0006] Incidentally, with the recent advancements in AI (Artificial Intelligence), conversation generation technology has been developed that allows computers to artificially generate conversations that are closer to human speech than ever before. Therefore, there has been a demand to reduce the workload on human operators by utilizing this conversation generation technology in call centers.

[0007] Therefore, while AI-powered technology is used to respond to inquiries by preparing preset answers in advance, there were problems with usability, such as the inability to select the appropriate preset answer or the inability to select the appropriate preset answer unless the answers were in the correct order.

[0008] This invention was made to solve these problems, and its purpose is to provide an easy-to-use answer generation device, answer generation program, answer generation method, and AI call center system. [Means for solving the problem]

[0009] To solve these problems, the AI ​​call center system of the present invention includes a CTI device that controls the line connection between the user terminal and the call center operator, A customer management device that manages customer information, including customer personal information and attributes, as well as usage history, by associating it with a customer ID (IDentification), A status management device that manages the current usage status of all customers, A response generation device, which is connected to the user terminal via the CTI device, generates a response voice signal that is a response to the user voice signal based on the user voice signal supplied from the user terminal as the call handler, the customer management device, and the current status management device, A switching determination device that determines whether or not a switch to a human operator is necessary based on the user voice signal and the response voice signal, The customer management device and the current status management device are connected to an operator terminal that is operated by a human operator. Equipped with, The CTI device is The system is characterized by switching the call handler to the operator terminal when the switching conditions are met. [Effects of the Invention]

[0010] This invention enables the realization of an easy-to-use response generation device, response generation program, response generation method, and AI call center system. [Brief explanation of the drawing]

[0011] [Figure 1] This is a schematic diagram showing the configuration of a conventional call center system. [Figure 2] This is a schematic diagram (1) for illustrating the AI ​​call center system in the first embodiment. [Figure 3] This is a block diagram showing the configuration of a CTI device. [Figure 4] This is a block diagram showing the configuration of the operator terminal. [Figure 5] This is a block diagram showing the configuration of the response generation device. [Figure 6] This is a block diagram showing the configuration of a customer management system. [Figure 7] This is a block diagram showing the configuration of the current status management device. [Figure 8] This is a simplified diagram illustrating the response generation process. [Figure 9] This is a simplified diagram used to explain the scenario. [Figure 10] This is a simplified diagram (1) for explaining the tasks and subtasks. [Figure 11] This is a simplified diagram (2) for explaining the tasks and subtasks. [Figure 12] This is a schematic diagram (2) for illustrating the AI ​​call center system in the first embodiment. [Figure 13] This is a flowchart illustrating the response generation process. [Figure 14] This is a flowchart illustrating the raw response generation process. [Figure 15]This is a schematic diagram (1) for describing the AI call center system in the second embodiment. [Figure 16] This is a block diagram showing the configuration of a conversation monitoring terminal. [Figure 17] This is a schematic diagram for describing the conversation monitoring screen (1). [Figure 18] This is a schematic diagram for describing the conversation monitoring screen (2). [Figure 19] This is a schematic diagram (2) for describing the AI call center system in the second embodiment. [Figure 20] This is a schematic diagram (3) for describing the AI call center system in the second embodiment. [Figure 21] This is a schematic diagram (4) for describing the AI call center system in the second embodiment. [Figure 22] This is a schematic diagram for describing the callback processing. [Modes for Carrying Out the Invention]

[0012] Hereinafter, modes for carrying out the present invention will be described with reference to the drawings.

[0013] <First Embodiment> Reference numeral 101 shown in Figure 2 indicates the AI call center system as a whole. In the AI call center system 101, a CTI device 2, a manned call center 3 having human operators, and an AI operator 5 are connected via a telecommunication circuit such as a wireless or wired Internet. Further, the manned call center 3 and the AI operator 5 are connected to a customer management device 7 and a status management device 8 via a telecommunication circuit.

[0014] As shown in Figure 3, the CTI device 2 is configured such that a control unit 21, consisting of a CPU (Central Processing Unit), ROM (Read Only Memory), and RAM (Random Access Memory), comprehensively controls the entire CTI device 2 and executes AI call center processing according to an AI call center program pre-stored in the ROM.

[0015] The CTI device 2 accepts all calls from the user terminal 4 and relays them between the operator terminal 3A or the AI ​​operator 5. The storage unit 26 stores a conversation log database (hereinafter, the database is abbreviated as DB) 81.

[0016] This conversation log DB88 registers the conversation log with the user as audio signals and text information of the conversation, with the conversation associated with a call ID that identifies the call. Although not shown in the diagram, the storage unit 26 may be an internal storage device or a large-capacity external storage device connected via the external interface 23.

[0017] This CTI device 2 may be an in-house server installed within the company, or it may be a so-called cloud server installed externally, or an appropriate combination of in-house servers and cloud servers may be used.

[0018] The manned call center 3 has multiple operator terminals 3A, and the CTI device 2 selects and connects to the operator terminal 3A to be connected for communication with the user.

[0019] As shown in Figure 4, the operator terminal 3A is controlled by a control unit 31 consisting of a CPU, ROM, and RAM, which comprehensively controls the entire operator terminal 3A and executes AI call center processing according to an AI call center program pre-stored in ROM.

[0020] As shown in Figure 5, the AI ​​operator 5 is controlled by a control unit 51 consisting of a CPU, ROM, and RAM, which comprehensively controls the entire AI operator 5 and executes AI call center processing according to an AI call center program pre-stored in ROM. Although not shown, the AI ​​operator 5 may be a single computer or may consist of multiple computers.

[0021] The storage unit 56 stores a conversation log database (hereinafter referred to as DB) 82. Similar to the conversation log DB 88, this conversation log DB 89 registers the conversation log with the user as audio signals of the call and text information of the call, associated with a call ID that identifies the call. Although not shown in the figures, the storage unit 56 may be an internal storage device or a large-capacity external storage device connected via the external interface 53.

[0022] As shown in Figure 6, the customer management device 7 is configured such that a control unit 71, consisting of a CPU, ROM, and RAM, comprehensively controls the entire customer management device 7 and executes various processes according to a program pre-stored in the ROM.

[0023] As shown in Figure 7, the customer management device 7 stores a customer management database in its storage unit 75, which manages personal information such as the customer's name and address, attribute information such as age, gender, and occupation, and the customer's past usage history with the service provider. The customer management device 7 may be part of the CTI device 2 as part of a cloud CTI system, for example, but for information management purposes, a server owned by the service provider or a cloud server used by that company is preferably used.

[0024] When the customer management device 7 receives an information request signal and the requested information from the operator terminal 3A or the AI ​​operator 5, it supplies the information represented by the information request signal to the operator terminal 3A or the response generation device. Furthermore, when the customer management device 7 receives an information registration signal and the information to be registered from the operator terminal 3A or the AI ​​operator 5, it supplies the information represented by the information request signal to the operator terminal 3A or the response generation device.

[0025] The current status management device 8 is configured such that a control unit 81, consisting of a CPU, ROM, and RAM, comprehensively controls the entire device and executes various processes according to a program pre-stored in the ROM.

[0026] The status management device 8, which manages the status of services and products provided by the providing company, can preferably be a server managed by the providing company or a server of a portal site company that provides a reservation management system for a portal site.

[0027] When the current status management device 8 receives an information request signal and the requested information from the operator terminal 3A or the AI ​​operator 5, it supplies the information represented by the information request signal to the operator terminal 3A or the response generation device. Furthermore, when the current status management device 8 receives an information registration signal and the information to be registered from the operator terminal 3A or the AI ​​operator 5, it supplies the information represented by the information request signal to the operator terminal 3A or the response generation device.

[0028] The AI ​​call center system 101 is designed to handle calls from user terminals 4 and receive various requests from users to service providers, such as orders, reservations, reservation changes, and repair requests. There are no restrictions on the services, products, or requests offered, but this explanation will focus on the case where the service provider sells products (mail order) and the CTI device 2 is a cloud server.

[0029] As shown in Figure 2, when user terminal 4 makes a call to a designated telephone number, the communication from user terminal 4 is connected to CTI device 2. In response to the call, CTI device 2 connects to AI operator 5, transmits information based on the dialed telephone number to AI operator 5, and relays the transmission and reception of spoken voice signals based on the user's utterances between user terminal 4 and AI operator 5.

[0030] As shown in Figure 8, when the control unit 51 of the AI ​​operator 5 receives a speech voice signal from the CTI device 2, the speech section detection unit 51A extracts the speech signal of the portion that was spoken and supplies it to the speech recognition unit 51B as the extracted speech signal.

[0031] The speech recognition unit 51B converts the extracted speech signal into text through speech recognition to generate spoken text, which it then supplies to the response generation unit 51C. The response generation unit 51C acquires information from the customer management device 7 and the current status management device 8, and then performs a response text generation process (details of which will be described later) in response to the spoken speech signal, generates the response text, and supplies the response text to the speech synthesis unit 51D.

[0032] The speech synthesis unit 51D synthesizes speech based on the response text and supplies it to the CTI device 2 as a response speech signal. In this way, talk processing is performed between the user and the AI ​​operator 5.

[0033] The memory unit 55 stores the conversation log DB, and utterance text and response text information are supplied from the response generation unit 51C and registered. In addition, the utterance voice signal supplied from the CTI device 2 and the response voice signal supplied from the speech synthesis unit 51D are registered in the conversation log DB 89.

[0034] Similarly, the CTI device 2 receives response text from the speech recognition unit 51B and the response generation unit 51C and registers it in the conversation log DB 88 stored in the storage unit 25. In addition, the voice signals supplied from the user terminal 4 and the speech synthesis unit 51D are also registered in the conversation log DB 89. These utterance texts, response texts, utterance voice signals, and response voice signals are stored in association with a call ID that identifies a single call from the start to the end of the call as a single unit.

[0035] As a result, conversation log DBs 88 and 89 record all audio signals and text information representing those audio signals exchanged between the user terminal 4 and the AI ​​operator 5, with the received ID associated with each signal.

[0036] The CTI device 2 supplies the response voice signal to the user terminal 4 (Figure 2). As a result, a conversation can take place between the user and the AI ​​operator 5.

[0037] Next, the response text generation process will be described. In this embodiment, the AI ​​call center system 101 is operated by a call center operating company, and the CTI device 2 is assumed to handle calls made to multiple companies.

[0038] As shown in Figure 8, the response generation unit 51C includes an AEI (Artificial Elastic Intelligence) 58 and an LLM (Large Language Models) 59. The AEI 58 is a response generation technology that incorporates machine learning and scenario-based methods, and is described, for example, in Japanese Patent No. 6951004.

[0039] As shown in Figure 9, the memory unit 55 has a scenario 90 registered for each providing company. For example, Company A has registered one scenario 91 for one phone number. Company B has registered a scenario 92 that branches into three for one phone number. Scenario 92 is a scenario that first checks the requirements and then branches into three.

[0040] Company C has registered two phone numbers and two scenarios, 93 and 94, associated with each phone number. Scenarios 91-94 have multiple tasks T and their sequence registered.

[0041] Specifically, when the CTI device 2 receives a call, it assigns a call ID, which is an identifier for each call, and supplies it to the AI ​​operator 5 along with the call destination telephone number, which is the telephone number specified by the user, and the voice signal. The AI ​​operator 5 supplies the speech text, which is based on the user's utterance generated by the speech segment detection unit 51A and the speech recognition unit 51B, to the response generation unit 51C along with the call destination telephone number and the call ID.

[0042] The AEI 58 of the response generation unit 51C selects a scenario associated with the recipient's telephone number. Here, we will explain the case where scenario 91 is selected.

[0043] As shown in Figure 10, Scenario 91 has tasks T1 to T5 registered, and the processing proceeds in the order of tasks T1, T2, T3, T4, and T5.

[0044] Each task T has one or more subtasks TS that represent specific instructions. Each subtask TS (subtasks T1-S1~T1~S2 set for task T1 in the diagram) has one or more preset answers registered.

[0045] For example, when the system first connects to user terminal 4, AEI 58 executes the first subtask T1-S1 "Standard greeting and asking for name," which is set in task T1 "Identifying customer ID." AEI 58 supplies the standard phrase "Thank you for calling. This is Tomato, the online shop. First, could you please tell me your name?" along with the incoming call ID as the response text to the speech synthesis unit 51D.

[0046] The speech synthesis unit 51D synthesizes speech based on the response text and supplies the synthesized response speech signal along with the call ID to the CTI device 2. The CTI device 2 transmits the response speech signal to the user terminal 4 indicated by the call ID. As a result, the user terminal 4 outputs speech based on the response speech signal.

[0047] When a user says, for example, "This is Taro Patent," the user terminal 4 supplies the voice collected by the microphone (not shown) as a spoken voice signal to the CTI device 2.

[0048] The spoken audio signal is supplied to the response generation unit 51C as the spoken text "Tokkyo Tarou desu" via the speech segment detection unit 51A and the speech recognition unit 51B.

[0049] The AEI 58 of the response generation unit 51C analyzes the meaning of the spoken text based on the syntactic analysis of the text and the preceding conversational exchange.

[0050] In other words, when AEI58 recognizes "Tokkyotaro" as the user's name, it executes a raw response generation process (details of which will be described later). Here, we will explain the case where one or more appropriate preset responses are selected from the preset responses registered in subtask T1-S1.

[0051] A preset response is a combination of semantic elements that represent meaning, such as subject, verb, and object, and these elements determine the meaning of the response. Furthermore, AEI58 selects the user's name, "Tokkyotaro," as a content element that the user wants to request, such as the user's name and order details. AEI58 uses the selected preset response and the user's name as the raw response and supplies it to LLM59 along with the utterance text and the instruction request that represents the sentence adjustment.

[0052] This instruction request is registered in the memory unit 55 and linked to a preset response. For example, in response to a preset response such as "Customer, Name = XX?", an instruction request is supplied saying, "Please replace XX with a name and adjust the sentence."

[0053] Here, LLM59 is a large-scale language model based on deep learning. Based on the utterance text, raw response, and request for instructions, LLM59 inserts the instructed words into the designated locations according to the request for instructions, and also performs sentence adjustments based on the raw response, such as adjusting conjunctions, particles, introductory phrases, and sentence endings, and sends the resulting draft response text to AEI58.

[0054] When a draft response text is provided, AEI58 analyzes its meaning and performs a sentence check to ensure that it does not deviate in meaning from the spoken text and the raw response. If it determines that there is a semantic discrepancy, it sends the same instruction request to LLM59 again.

[0055] If the same instruction request is submitted again, LLM59 will reconsider its meaning, revise the text again, and send a different draft response text to AEI59.

[0056] If the AEI58 determines that the proposed response text is appropriate after checking the text, it supplies the proposed response text to the speech synthesis unit 51D as the response text. As a result, the response audio signal is supplied to the CTI device 2, and the user terminal 4 outputs audio based on the response audio signal.

[0057] As shown in Figure 10, the AI ​​operator 5 engages in conversation with the user in the same manner and completes the first task, Task T1. Task T1 is a task to identify the user. Through interaction with the user, the AI ​​operator 5 obtains information such as the user's name and phone number, name and address, and customer ID, and then compares this information with the customer management database stored in the customer management device 7 to identify the customer ID.

[0058] The customer management database contains user attribute information such as name, address, gender, age, occupation, and residential area, as well as past purchase history. When a customer ID is identified, AEI58 queries the customer management database to retrieve the user's attribute information.

[0059] Therefore, in Task T2 and beyond, AI Operator 5 requests LLM59 to generate appropriate response text suggestions based on the user's attribute information. This enables AI Operator 5 to create appropriate response text tailored to the user.

[0060] Task T2 is for order acceptance, and Task T3 is for delivery address verification. If the address has changed, the system will ask the user to confirm the new address and preferred delivery date. In Task T4, as a closing step, the system will confirm the order details, delivery address, and estimated delivery date with the user.

[0061] In the final task, T5, post-call processing and integration processing are performed after the call has ended. For example, this includes confirming and registering order details and address changes, registering the end of the conversation, and sending order information via SMS (Short Message System) or registered email addresses.

[0062] As shown in Figure 11, subtasks T2-S1 to T2-S2 of task T-2 are cyclical and repeat until the spoken text indicates the end. Specifically, in subtask T2-S1, an order is taken, and based on the spoken text, the current status management DB registered in the current status management device 8 is queried to confirm the inventory of the ordered goods.

[0063] In subtask T2-S2, if there is no stock or insufficient stock, AEI58 selects a preset response to that effect as the basic response. If there is stock, it repeats the preset response from subtask T2-S1 without providing any response. On the other hand, if the user asks about the availability of stock, subtask T2-S2 communicates the stock status.

[0064] Furthermore, when AEI58 receives speech text based on user utterances, it analyzes the content of the response and registers the necessary information in the customer management device 7 or the current status management device 8. AEI58 also generates preset response sentences for the user from previous conversations with the user and supplies them to LLM59.

[0065] LLM59 adjusts the preset response text supplied by AEI58 to match the user's needs.

[0066] In other words, AEI58 determines the dynamic elements of the conversation, such as the subject, predicate, and object, which constitute the content of the response, generates multiple preset response sentences, and supplies them to LLM59. Based on the user's attributes, LLM59 adjusts the response text sentences provided by AEI58 to match the speaking style that the user is likely to prefer.

[0067] This allows AI Operator 5 to adjust its speaking style to suit the user, thereby increasing user satisfaction.

[0068] Furthermore, if AI Operator 5 determines that it cannot select an appropriate preset response because the spoken text (user's response) does not conform to the subtask TS, it will search for preset responses in the following order: another subtask TS within the same task T, and then another task T.

[0069] For example, in the case of task T2 in Figure 11, subtask T2-S1 takes an order from the user. AEI58 selects, for example, "I will take your order." or "Please place your order." and supplies it to LLM59 along with the call ID and an instruction request indicating the wording adjustment.

[0070] LLM59 stores the conversation history and user attribute information provided for the incoming call ID. Based on the past conversation and user attribute information, it adjusts the text and supplies a draft response text to AEI58.

[0071] When AEI58 receives the utterance text "I'd like one tomato juice and two apple juices, please," it recognizes that an order has been placed and executes the processing of subtask T2-S2. AEI58 recognizes that this means 1 tomato juice and 2 apple juices, and requests the current status management device 8 to transmit information on the stock quantities of tomato juice and apple juice.

[0072] The current status management device 8 queries the current status management database and supplies the current stock quantities of tomato juice and apple juice to the AEI 58. If there is a problem with the stock, the AEI 58 selects a preset response message to that effect, confirms the draft response text adjusted by the LLM 59, and then sends it to the CTI device 2.

[0073] On the other hand, if AEI58 determines that there are no problems with inventory, it selects a preset response for subtask T2-S4 and generates a response text indicating whether or not an additional order is to be placed.

[0074] In other words, AEI58, based on the generated text, first searches for the next subtask (T2-S3) following the previously processed subtask (T2-S2), looking for an appropriate preset response or processing based on the response text.

[0075] If no appropriate preset answer or action is found for the next task, then the system searches for other subtasks within the same task (T2) (T2-S2, T2-S1, T2-S4) for appropriate preset answers or actions corresponding to the answer text.

[0076] Note that the last subtask of each task T (in this case, T2-S4) is the final subtask within the task that will not be executed unless the processing of the preceding subtask (in this case, T2-S4) has been completed. It is flagged as a conditional task with conditions attached to its execution, and is excluded from the search if the preceding processing is not completed.

[0077] In other words, AEI58 sets a completion flag for subtasks TS in a scenario that have been completed. If there is no appropriate preset answer or action for the next subtask TS based on the answer text, it first searches for incomplete subtasks TS within the same task T, and then searches for completed subtasks TS within the same task T.

[0078] This allows AI operator 5 to perform appropriate processing or supply response text even when orders are repeated in the response text.

[0079] Furthermore, if AEI58 cannot find a suitable preset answer or solution within the same task (T2), it will also search other tasks (T3, T4, T1). Note that the last task (T5 in this case) is the final task and will not be executed until the processing of the preceding task (T4 in this case) is complete, so it is excluded from the search.

[0080] AEI58 flags completed tasks T in a scenario as complete. If there is no appropriate preset answer or action for the next task T based on the answer text, it first searches for incomplete tasks T, and then searches for completed tasks T.

[0081] For example, once the inventory check is complete, AEI58 moves to task T2-S4 and generates a response text asking for an additional order. After this, when an utterance text indicating that an additional order will be placed is supplied, there is no appropriate action or preset response in the next subtask T2-S5. Therefore, AEI58 searches within the same task (T2) and executes the appropriate subtask (T2-S2).

[0082] This allows AI Operator 5 to always select and execute an appropriate preset response or action, even when unexpected timings or content are provided, such as when the content of the response text changes or when the user suddenly says something different. Furthermore, because it can sequentially search from the most likely subtasks TS and task T, it is possible to minimize the search processing time.

[0083] Furthermore, if, despite being asked for an order in subtask T2-S1, the response text is "Actually, I've moved," and there is no appropriate processing or preset response in task T2, the system searches for the next task T3 (checking the delivery address), executes task T3 first to complete the process, and then executes task T2.

[0084] This allows AI Operator 5 to respond even if the user gives answers that are out of order, and to conduct conversations at the user's pace.

[0085] In other words, AI operator 5 configures a task T to be executed and a subtask TS that represents a preset response or process that describes the detailed conversation content. The subtasks are executed in order, and when one task is completed, the next task is executed.

[0086] At this time, if there is no suitable preset answer or action in the next subtask TS, the AI ​​operator 5 searches within task T, and if there is no suitable preset answer or action within task T, it searches other tasks T.

[0087] In other words, AEI58 generally proceeds with a conversation in the order set by default for several pre-configured tasks T, but it is possible to change the order of tasks T depending on the spoken text.

[0088] This allows AI operator 5 to respond flexibly even if there are cyclical subtasks TS or if the order of tasks T changes.

[0089] Furthermore, the AI ​​call center system 101 of the present invention includes a manned call center 3, and if it is determined that there is no appropriate preset response or processing within task T, it switches to the manned call center 3.

[0090] Specifically, as shown in Figure 12, AEI58 determines that a scenario has been deviated from if it searches for processing or preset answers for all tasks except conditional tasks but finds no suitable results.

[0091] AEI58 requests LLM59 to create a summary of the conversation so far, and then transmits the created summary to CTI device 2 along with a switch request signal to manned call center 3 and the call ID.

[0092] When a switching request signal is supplied, the control unit 21 of the CTI device 2 checks the call status of the manned call center 3 and switches the destination of the spoken voice (i.e., the user's call recipient) to an available operator terminal 3A. At this time, the control unit 21 transmits all spoken text and response text, summary text, and customer ID associated with the call ID to the operator terminal 3A.

[0093] The control unit 31 of the operator terminal 3A displays the supplied conversation history (spoken text and response text) and summary on the display unit 34 as soon as a call is connected. The control unit 31 also transmits the customer ID and an information request signal to the customer management device 7, retrieves customer information from the customer management device 7, and displays it on the display unit 34. Therefore, the human operator can quickly grasp the conversation history and customer information and continue the conversation with the user without repeating the previous conversation.

[0094] Next, the response generation process RT1, which is executed according to the response generation program, will be explained using the system chart shown in Figure 13.

[0095] In step S11, when the AEI 58 of the response generation unit 51C receives the call ID and spoken text from the speech recognition unit 51B, it moves to the next step S12. In step S12, the AEI 58 recognizes the meaning of the spoken text while considering its relationship to the preceding and following conversational sentences corresponding to the call ID, and moves to the next step S13.

[0096] In step S13, AEI58 executes the raw response generation process (subroutine SRT11) and moves to the next step S14. In step S14, AEI58 sends the raw response and the request for text adjustment to LLM59.

[0097] In step S16, LLM59 adjusts the raw response text and moves to the next step S17. In step S17, LLM59 sends the adjusted response text, the draft response text, to AEI58.

[0098] In step S18, AEI58 checks the raw response text and moves on to the next step, S19. In step S19,

[0099] Next, the raw response search procedure SRT11, which is executed in step S13 of the raw response generation procedure RT1 (Figure 13), will be explained using the flowchart in Figure 14.

[0100] AEI58 starts the raw response search process from step S13 of the raw response generation process procedure RT1 and moves to step S21. In step S21, when the response text is supplied, AEI58 searches for the current subtask TS (the one following the previously completed subtask TS).

[0101] In step S22, AEI58 determines whether the appropriate preset answer was selected. If a negative result is obtained here, it means that the user may be talking about something else, and in this case AEI58 moves on to the next step S23.

[0102] In step S23, AEI58 searches for other subtasks TS of task T to which the current subtask TS belongs that have not yet been processed.

[0103] In step S24, AEI58 determines whether the appropriate preset answer was selected. If a negative result is obtained here, it means that the user may be talking about something else, and in this case AEI58 moves on to the next step S25.

[0104] In step S25, AEI58 searches for other subtasks TS of task T to which the current subtask TS belongs, and for which processing has been completed.

[0105] In step S26, AEI58 determines whether the appropriate preset answer was selected. If a negative result is obtained here, it means that the user may be talking about something else, and in this case AEI58 moves on to the next step S27.

[0106] In step S27, AEI58 searches for tasks T other than the current task T that are not yet completed.

[0107] In step S28, AEI58 determines whether the appropriate preset answer was selected. If a negative result is obtained here, it means that the user may be talking about something else, and in this case AEI58 moves on to the next step S29.

[0108] In step S29, AEI58 searches for tasks T other than the current task T that have already completed processing.

[0109] In step S300, AEI58 determines whether it was able to select an appropriate preset answer. If a negative result is obtained here, this means that it was likely the user was talking about something other than task T, since it was unable to select anything despite searching all possible tasks T, and in this case AEI58 moves on to the next step S37.

[0110] In step S37, AEI58 determines that the user is talking about something outside the scenario and proceeds to the next step S38.

[0111] In step S38, AEI58 requests the CTI device 2 to switch to the manned call center 3 and terminates the raw response generation process.

[0112] In contrast, if positive results are obtained in steps S22, S24, S26, S28, and S30, this indicates that an appropriate preset answer has been selected and no further searching is necessary. At this point, AEI58 proceeds to the next step, S39.

[0113] In step S39, AEI58 determines the selected preset answer as the raw answer and proceeds from step S13 (Figure 13) to step S14 of the raw answer generation process RT1.

[0114] <Second Embodiment> Next, a second embodiment will be described using Figures 15 to 20. The second embodiment differs from the first embodiment in that it has a conversation monitoring device 209 that monitors the content of the conversation between the AI ​​operator 205 and the user, and the AI ​​operator 205 performs callback reservation processing. In the second embodiment, the same reference numerals are used for the same parts as in the first embodiment, and 200 is added to the reference numerals for corresponding parts, and the explanation of the same parts is omitted.

[0115] In the AI ​​call center system 201 of the second embodiment, the CTI device 202 relays the transmission and reception of voice signals between the user terminal 4 and the AI ​​operator 205, similar to the first embodiment.

[0116] As shown in Figure 15, the AI ​​call center system 201 includes a conversation monitoring device 209. In this embodiment, the AI ​​call center system 201 will be described in the case where it accepts airline ticket reservations and changes.

[0117] As shown in Figure 16, the conversation monitoring device 209 is controlled by a control unit 291 consisting of a CPU, ROM, and RAM, which comprehensively controls the entire conversation monitoring device 209 and executes AI call center processing according to an AI call center program pre-stored in ROM.

[0118] In the AI ​​call center system 201, a conversation takes place between the CTI device 202 and the AI ​​operator 205 in the same manner as in the first embodiment, and the voice text (spoken text and response text) is registered in the conversation log DB 88 of the CTI device 202.

[0119] Furthermore, in the AI ​​call center system 201, voice-to-text is monitored by a monitoring operator via a conversation monitoring device 209.

[0120] As shown in Figure 17, the conversation monitoring device 209 displays the conversation monitoring screen 270X on the display unit 295. Here, information about the user "Tataku Taro" is displayed. The conversation monitoring screen 270 displays a voice text display field 271, a user information display field 272, an other user display field 273, and a toggle button 274.

[0121] The voice-to-text display area 271 displays the voice-to-text of the conversation history. The voice-to-text is displayed using different colors, fonts, and formatting for the spoken text and the response text so that it can be easily distinguished at a glance whether it is the user's spoken text or the AI ​​operator's response text. In the diagram, an example is shown where the spoken text is underlined and the response text is italicized.

[0122] The user information display area 272 displays user personal information such as the user's name, user attribute information, and the details of the airline ticket reservation corresponding to the reservation number. By simply looking at the voice-text display area 271 and the user information display area 272, the monitoring operator can recognize who made which reservation and what kind of conversation took place between the AI ​​operator 205 and the user.

[0123] Here, it is assumed that each monitoring operator will monitor multiple conversations. The other user display field 273 shows the names of the other users that the monitoring operator is dealing with. The switch button 274 is a button that requests a switch to the manned call center 203.

[0124] Here, the control unit 291 of the conversation monitoring device 209 monitors the supplied voice text, and if it is determined that the conversation is unnatural, such as when NG words like "wrong" or "that's not right" are detected two or more times in the utterance text, or when sentences with the same meaning are detected three or more times consecutively in the response text, it determines that a switching factor has been detected and notifies the monitoring operator of an alert.

[0125] In Figure 17, an alert notification 276 pops up for "Trademark Jiro" among other users 273A, and "Trademark Jiro" in the other users display section 273 is flashing to stand out. Currently, the conversation monitoring screen 270X for "Patent Taro" is displayed, so operating the switch button 274 will not trigger a switch request.

[0126] If another user 273A operates "Trademark Jiro," the conversation monitoring screen 270Y for "Trademark Jiro" is displayed, as shown in Figure 18. On the conversation monitoring screen 270Y, the toggle button 274 is highlighted to stand out by changing color or flashing.

[0127] When the switch button 274 is operated in this state, a switch request process is executed on the CTI device 202, as shown in Figure 19. The CTI device 202 checks the congestion status of the manned call center 203 and sends a call reception ID and a callback instruction signal to the AI ​​operator 205 to perform a switch process if it can be switched immediately, or to perform a callback reservation process to reserve a callback if it determines that switching is difficult.

[0128] Specifically, the storage unit 225 of the CTI device 202 stores a call management list, which manages the current call status of operator terminal 3A in the manned call center 203. This call management list is updated at predetermined intervals (for example, every second), so that it always reflects the latest status.

[0129] The control unit 291 of the conversation monitoring device 209 transmits the power reception ID and a switching request signal to the CTI device 202 as part of the switching request processing.

[0130] When a switching request signal is supplied, the control unit 221 of the CTI device 202 checks the call management list and, if it determines that it should connect to the manned call center 203, executes the switching process.

[0131] The control unit 221 transmits the power reception ID and a switching instruction signal to the AI ​​operator 205. As shown in Figure 20, the AI ​​operator 205 supplies an audio signal to the CTI device 202 via the CTI device 202, which represents a switching announcement such as, "We will switch you to a human operator, please wait a moment."

[0132] The control unit 221 of the CTI device 202 transmits a voice signal for switching guidance to the user terminal 4, and then switches to the operator terminal 3A of the manned call center 3. As a result, a conversation takes place between the human operator and the user via the operator terminal 3A, the CTI device 202, and the user terminal 4.

[0133] In response, when the control unit 221 of the CTI device 202 receives a switching request signal and checks the call management list, if it determines that it should not connect to the manned call center 203, it supplies a callback instruction signal to the AI ​​operator 205 and executes the callback reservation process (see Figure 21).

[0134] Furthermore, on the conversation monitoring screen 270Y, the answer input button 275 is displayed, highlighted like the toggle button, and is in an operable state. When the monitoring operator operates the answer input button 275, an answer stop request signal is supplied to the answer generation unit 51C via the CTI device 2, and a blank space for input is displayed on the conversation monitoring screen 270 (not shown).

[0135] When a monitoring operator enters a response to a blank, such as "Has your mother already reserved tickets?", the monitoring response entered by the monitoring operator is supplied to the response generation unit 51C via the CTI device 2.

[0136] The response generation unit 51C generates response text based on the monitored response, and as a result, the user terminal 4 receives a response audio signal based on the monitored response. Subsequently, if the conversation with the AI ​​operator 5 proceeds smoothly, the alert disappears and the system returns to normal mode.

[0137] In this way, if the conversation between AI Operator 5 and the user becomes stalled, a human monitoring operator can intervene to restore the flow, allowing AI Operator 5 to handle most of the phone calls.

[0138] Figure 22 shows scenario 290 in the second embodiment.

[0139] Specifically, in task T101, the control unit 221 of the CTI device 202 identifies the request summary, verifies the name in task T111 if the request is for an already booked airline ticket, and identifies the reservation number in task T102 if the request concerns an already booked airline ticket.

[0140] Task T103 verifies the user's identity based on the name associated with the reservation number, and Task T104 verifies the flight associated with the reservation number. Task T106 gathers detailed requests from the user, and Task T107 executes the requested changes or cancellations, followed by closing in Task T4 and post-processing / integration in Task T5.

[0141] In the AI ​​call center system 201, the AI ​​operator 205 makes a decision based on the request summary (task T101) or the request details (task T106) and executes a switching request process to the CTI device 2.

[0142] In other words, if the AI ​​operator 205's AEI 258 determines that it should switch to the manned call center 203 in response to the user's request, it moves to task T105 and transmits the incoming call ID and a switching request signal to the CTI device 202.

[0143] When a switching request signal is supplied, the control unit 221 of the CTI device 202 checks the call management list and, if it determines that it should connect to the manned call center 203, executes the switching process.

[0144] The control unit 221 transmits the power reception ID and a switching instruction signal to the AI ​​operator 205. As shown in Figure 20, the AI ​​operator 205 supplies an audio signal to the CTI device 202 via the CTI device 202, which represents a switching announcement such as, "We will switch you to a human operator, please wait a moment."

[0145] In response, when the control unit 221 of the CTI device 202 checks the call management list and determines that it should not connect to the manned call center 203, it sends the incoming call ID and a callback instruction signal to the AI ​​operator 205 to execute the action.

[0146] When the AI ​​operator 205's AEI 258 receives a loopback instruction signal from the CTI device 202, it performs the loopback guidance process according to task T109.

[0147] Specifically, AEI258 determines the priority of the callback based on the previous conversation and selects a preset response for subtask T109-S1. For example, if the flight change is within two hours, it checks the CTI device 202 for the fastest estimated callback time, selects the preset response corresponding to priority 1, and communicates the estimated time.

[0148] For flight changes on the same day or the following day or later, the system prioritizes them second, while simple inquiries or questions are given the lowest priority, and the system obtains a date and time convenient for the user. For example, in the case of a second-priority request, the system sends a message to CTI device 202, checks the congestion status of the return list for that day, and then selects a preset response.

[0149] For example, if the priority is the lowest and the callback list for that day is congested, a preset response asking about the user's availability for the following day or later will be selected. Regarding priority, a priority table with rules for assigning priority is registered in the memory unit 255, and priority is determined according to the importance of the content and the schedule such as the reservation date and time.

[0150] <Action and Effects> The following describes the features of the group of inventions extracted from the embodiments described above, showing effects and other details as necessary. For ease of understanding, the corresponding configurations in each embodiment will be indicated in parentheses as appropriate, but the invention is not limited to these specific configurations. Furthermore, the meanings and examples of terms described in each feature may also be applied to the meanings and examples of terms described in other features that use the same wording.

[0151] <Feature A-1> According to the above configuration, the AI ​​call center system of the present invention includes a CTI device (CTI device 2) that controls the line connection between the user terminal (user terminal 4) and the call operator, A customer management device (customer management device 7) manages customer information, including customer personal information and attributes, as well as usage history, by associating it with a customer ID (IDentification), A status management device (status management device 8) that manages the current usage status of all customers, A response generation device (AI operator 5) is connected to the user terminal via the CTI device and generates a response voice signal that is a response to the user voice signal based on the user voice signal supplied from the user terminal as the call operator, the customer management device, and the current status management device, Based on the user voice signal and the response voice signal, a switching determination device (AI operator 5, conversation monitoring device 209) determines whether or not it is necessary to switch to a human operator, The customer management device and the current status management device are connected to an operator terminal (operator terminal 3A) which is operated by a human operator. Equipped with, The CTI device is When the switching conditions are met, the call handler is switched to the operator terminal. Characterized by

[0152] This means that in an AI call center system, it is only necessary to switch to an operator terminal used by a human operator when the AI ​​operator is unable to handle the call, thus reducing the number of operators required.

[0153] <Feature A-2> In the AI ​​call center system described in Feature A-1, The aforementioned response generation device is As the call handler, the system generates spoken text from the user voice signal supplied from the user terminal, generates a response text based on the spoken text, the customer information, and the usage history information, and generates a response voice signal from the response text. The utterance text and the response text are transmitted to the CTI device. The CTI device is The aforementioned utterance text and the aforementioned response text are stored, The system is characterized by transmitting the spoken text and the response text to the operator terminal when switching to the operator terminal.

[0154] This allows the AI ​​call center system to supply the conversation history to the operator terminal when the AI ​​operator sends a message, enabling the operator to quickly understand the content of the conversation so far.

[0155] <Feature A-3> In the AI ​​call center system described in Feature A-2, the CTI device is Each call from the user terminal is associated with a call ID, and the user voice signal and the call ID are associated and supplied to the response generation device. The aforementioned response generation device is The customer ID is obtained from the exchange of the user voice signal and the response voice signal, and when switching to the operator terminal, the customer ID is supplied to the CTI device. The CTI device is When switching to the operator terminal, the customer ID is transmitted to the operator terminal. The aforementioned operator terminal is The customer information obtained from the customer management device based on the aforementioned customer ID, along with the spoken text and the response text, are displayed on the display unit.

[0156] This allows the AI ​​call center system to display customer information on the operator's terminal, enabling the AI ​​operator to fully utilize the information they have acquired so far.

[0157] Alternatively, the response generation device may create a conversation summary text of the spoken text and the response text and send it to the CTI device, and the CTI device may send the conversation summary text along with the customer ID to the operator terminal.

[0158] This means that in an AI call center system, operators only need to read a summary, allowing them to quickly grasp the conversation history.

[0159] <Feature A-4> In the AI ​​call center system described in Feature A3, the switching discrimination device is It is part of the aforementioned response generation device, The CTI device is When a switching request signal is supplied from the response generation device, The call handling is characterized by switching from the response generation device to the operator terminal.

[0160] This allows the AI ​​call center system to easily switch to an operator terminal based on the response generation device's judgment.

[0161] <Feature A-5> In the AI ​​call center system described in Feature A-3, The aforementioned switching discrimination device is This is a conversation monitoring device (conversation monitoring device 209) that displays the utterance text and the response text on a display unit and allows a monitoring operator to monitor them. The call handling role is switched to the operator terminal in response to the operation input of the monitoring operator to the conversation monitoring device.

[0162] This allows the AI ​​call center system to switch to a human (supervising operator) terminal only when the AI ​​operator is unable to handle the call, based on the operator's judgment.

[0163] <Feature A-6> In the AI ​​call center system described in Feature A-5, the conversation monitoring device is: The system is characterized in that, when any of the following are detected as switching factors: repetition of a negative word in the utterance text, repetition of an answer with the same meaning in the answer text, or determination that an appropriate answer cannot be given, the display unit displays a switching button indicating switching to a manned operator terminal and an alert indicating that a switching factor has been detected, and when the monitoring operator operates the switching button, the system determines that the switching conditions have been met and transmits a switching request signal to the CTI device.

[0164] This means that the AI ​​call center system only needs to have monitoring operators respond when an alert is displayed, thus reducing the number of monitoring operators required.

[0165] <Feature A-7> In the AI ​​call center system described in Feature A-7, the conversation monitoring device is: The system is characterized in that, when any of the following are detected as switching factors—repetition of negative words in the utterance text, repetition of answers with the same meaning in the response text, or determination that an appropriate answer cannot be given—the display unit is shown a blank space for character input and an alert indicating that a switching factor has been detected, and the text entered by the monitoring operator is supplied to the response generation device as the response text.

[0166] This allows the AI ​​call center system to have a human (supervising operator) handle the call only for a short period when a problem occurs, and then switch to an AI operator. Furthermore, since the user appears to be receiving the same operator (AI operator), they will not experience the discomfort that users might feel from a change in operator.

[0167] <Feature A-8> In the AI ​​call center system described in Feature A-2, the CTI device is It remembers the callback schedule, which represents the schedule for return calls. The aforementioned response generation device is When it is determined that the operator should switch to the aforementioned operator terminal, the CTI device checks the congestion status, The CTI device is Based on the current status of the operator terminal, the system requests the response generation device to begin the callback preparation process to inquire about the user's availability. The aforementioned response generation device is The system is characterized by setting priorities based on the conversation history and generating response text according to the said priorities.

[0168] This allows the AI ​​call center system to ask users about their availability before calling back during busy periods for human operators (operator terminals), enabling the creation of appropriate callback schedules based on the priority of each case.

[0169] <Feature A-9> In the AI ​​call center system described in Feature A-4, the response generation device is: Based on the meaning of the analyzed utterance text, a preset response corresponding to a pre-configured scenario is selected, response text is generated based on the selected preset response, and if no preset response can be selected, it is determined that the switching condition has been met. The AI ​​call center system according to feature 2.

[0170] This allows the response generation device to switch to the operator terminal if the spoken text is different from the pre-prepared responses, enabling a smooth transition.

[0171] <Feature B-1> In the above configuration, the answer generation device (AI operator 5) of the present invention generates answer text in response to user utterances supplied from a user terminal owned by the user, and engages in conversation with the user by supplying the answer based on the answer text to the user terminal, A storage unit (storage unit 55) stores multiple tasks to be performed as specified in a conversation with the user (task T), multiple subtasks (subtask TS) associated with the task and containing details for executing the task, the order in which the task and subtasks are executed, and multiple preset responses associated with the subtasks. The system includes a response generation unit (response generation unit 51C) that analyzes the meaning of the utterance text by referring to a conversation history consisting of utterance text and response text based on the user's utterances, selects an appropriate preset response in the utterance text, and generates a response text based on the selected preset response, It includes a response audio signal generation unit that synthesizes speech based on the response text and generates a response audio signal, The aforementioned response generation unit, The system searches for a preset answer associated with the currently running subtask based on the spoken text. If there is no suitable preset answer for the currently running subtask, it searches for a preset answer within the same task. If a suitable preset answer is not available within the same task, the system searches for a preset answer in another task.

[0172] This allows the answer generation device to perform searches from the most likely subtasks and tasks, thereby reducing the time required for searches and enabling the rapid generation of answer text.

[0173] <Feature B-2> In the response generation device described in Feature B-1, the response generation unit is: The system remembers whether the selection of preset answers for subtasks and tasks is complete or incomplete. If there is no suitable preset answer for the currently running subtask, the system will first search for incomplete preset answers within the same task, and then search for completed preset answers within the same task. If there is no suitable preset answer within the same task, search for preset answers in other incomplete tasks, and then search for preset answers in other completed tasks. The answer generation device according to feature 1.

[0174] This allows the answer generation device to perform searches from the most likely subtasks and tasks, thereby reducing the time required for searches and enabling the rapid generation of answer text.

[0175] <Feature B-3> In the response generation device described in Feature B-1, the response generation unit is: The system is characterized by searching for preset answers within all tasks to be searched, and performing error handling if a suitable answer cannot be selected. Error handling may include, for example, creating a callback schedule or switching to another terminal (an operator terminal or a dedicated error handling terminal).

[0176] This allows the response generation device to determine that the conversation is not what was expected if there is no preset response, and to perform error handling at the appropriate time.

[0177] <Feature B-4> In the response generation device described in Feature B-3, the response generation unit is: The error handling method is characterized by executing a request to switch to another device.

[0178] This allows the response generation device to quickly switch to another device if it determines that it cannot process the data itself.

[0179] <Feature B-5> The response generation device described in Feature B-1, The aforementioned utterance text and the aforementioned response text are displayed on the display unit (display unit 293), allowing the monitoring operator to monitor them. A response generation system (AI call center system) characterized by having a conversation monitoring device (conversation monitoring device 209) that, when any of the following are detected as trouble factors: repetition of negative words in the utterance text, repetition of responses with the same meaning in the response text, or determination that an appropriate response cannot be given, displays a blank space for character input and an alert indicating that a trouble factor has been detected on the display unit, and supplies the input text entered by the monitoring operator as the response text to the response generation device.

[0180] This allows a monitoring operator to temporarily intervene in the conversation between the user and the response generation device in the response generation system, resolving any communication blockages caused by problems and allowing the response generation device to continue its response.

[0181] <Feature B-6> In the response generation system described in Feature B-5, It has a storage unit that stores the utterance text and the response text, and the input text entered by the monitoring operator, Add preset answers based on the input text. The answer generation system according to claim 5, characterized by having the following features.

[0182] This allows the system to learn preset answers, potentially improving the accuracy of the answers generated by the answer generator.

[0183] <Feature C-1> According to the above configuration, the answer generation device of the present invention is an answer generation device (AI operator 5) that generates answer text in response to user utterances supplied from a user terminal owned by the user, and engages in conversation with the user by supplying the answer based on the answer text to the user terminal, The analysis unit (AEI58) analyzes the meaning based on the speech text derived from the user's utterances, A raw response generation unit (AEI58) generates raw responses that form the basis of answers to user utterances, consisting of semantic elements that represent meaning, such as subject, verb, and object, and content elements that the user wants to request, such as the user's name and order details. The Artificial Elastic Intelligence (AEI) unit (AEI58) has a document adjustment request unit that requests document adjustments, which is transmitted along with instructions for document adjustments. The system is characterized by having an LLM (Large Language Models) unit (LLM59) that performs sentence adjustments, such as adjusting conjunctions, particles, and introductory phrases, based on the aforementioned raw response, and supplies the draft response text to the AEI unit.

[0184] As a result, in the answer generation device of the present invention, the LLM unit can be used solely for text adjustment. Compared to a case where the LLM handles all conversational interactions, this eliminates the possibility of generating incorrect answers, which is a drawback of LLM, while reducing the processing load and enabling the generation of smooth, human-like answer text in a very short time.

[0185] <Feature C-2> In the response generation device described in Feature C-1, the analysis unit is: The method is characterized by analyzing the meaning of the provided draft response text and using it as the response text if it does not differ from the user's own intention (meaning-determining elements).

[0186] This allows the answer generation device of the present invention to verify that the meaning has not changed due to text adjustments before using the answer text.

[0187] <Feature C-3> In the response generation device described in Feature C-1, the analysis unit is: The system is characterized by analyzing the meaning of the provided response text draft, and if it is determined that it differs from the user's intention, modifying the content of the instruction request and transmitting the modified instruction content along with the semantic determination elements to the LLM unit.

[0188] As a result, in the answer generation device of the present invention, if the meaning changes due to text adjustment, the instruction content can be modified to prevent the same mistake from being repeated.

[0189] <Feature C-4> In the response generation device described in features C-1 to C-3, The aforementioned AEI is, It has a user attribute information acquisition unit that acquires user attribute information from an external customer management device based on the customer ID (Identification) obtained during conversations with the user. The aforementioned document adjustment request unit, The user attribute information is transmitted along with the instruction request and the semantic determination elements. The LLM unit performs text adjustments according to user attribute information. It is characterized by the following:

[0190] As a result, the answer generation device of the present invention can perform text adjustments according to user attributes, thereby creating answer text with appropriate wording for each user.

[0191] <Feature C-5> In the response generation device described in Feature C-4, The aforementioned semantic determining elements are preset answers with predetermined combinations. The aforementioned raw response generation unit selects a number of preset responses, The aforementioned text adjustment request unit sends multiple preset responses to the LLM. It is characterized by having the following features.

[0192] This allows the answer generation device to select semantic determination elements with simple processing, significantly reducing processing speed compared to when the answer is generated in the LLM unit.

[0193] <Other Embodiments> In the above embodiment, the conversation monitoring device 209 determines that a switching factor has been detected when an NG word is detected and notifies an alert, but the present invention is not limited to this. For example, even if the AEI determines that it is outside the scenario, it may be determined that a switching factor has occurred and an alert may be notified to the conversation monitoring device 209. In this case, it is preferable that the AI ​​operator does not make a switching request, but rather the conversation monitoring device 209 makes the switching request.

[0194] In the above embodiment, the customer management DB and the status management DB were described as being separate entities, but the present invention is not limited to this, and they may be managed as a single DB. Furthermore, the customer management device 7 and the status management device 8 may be a single device.

[0195] In the above embodiment, we described a case where AEI 58 and LLM 59 are used in combination as the answer generation unit 51C, but the present invention is not limited to this, and AEI alone may be used. In this case, it is preferable to enable the generation of answer text simply by selecting a preset answer, as in the second embodiment.

[0196] In the above embodiment, we described a case where scenario 90 has a structure using tasks and subtasks, but the present invention is not limited thereto. For example, preset answers may be registered for each related keyword, and searches may be performed starting with those closest to the keyword.

[0197] The above embodiment describes a case where raw answers are generated by selecting preset answers, but the present invention is not limited to this. For example, subjects, verbs, and objects may be assigned to subtasks separately, and the AEI may appropriately select these to generate raw answers.

[0198] The above-described embodiment described the application of the present invention to a call center that communicates with users via voice calls, but the present invention is not limited to this. For example, the present invention can also be applied to chatbots that communicate with users via text. [Industrial applicability]

[0199] This invention can be applied, for example, to a call center that receives inquiries by telephone. [Explanation of symbols]

[0200] 3: Manned call center 3A: Operator terminal 4: User terminal 5: AI Operator 7:Customer management device 8:Current situation management device 21: Control Unit 23: External Interface 25: Storage section 26: Storage and Memory Unit 31: Control Unit 34:Display section 51: Control Unit 51A: Speech segment detection unit 51B: Speech Recognition Unit 51C: Answer generation part 51D: Speech synthesis unit 53: External Interface 55: Storage section 56: Storage and Memory Unit 58: AEI 59: LLM 89: Conversation Log 270Y: Conversation monitoring screen 271: Audio-text display field 272: User information display field 273: Other users' display field 88: Conversation Log Database 89: Conversation Log Database

Claims

1. A response generation device that generates response text in response to user utterances supplied from a user terminal owned by the user, and engages in conversation with the user by supplying the response based on said response text to the user terminal, A storage unit that stores multiple tasks to be performed as specified in a conversation with the user, multiple subtasks associated with those tasks and which are details for performing those tasks, the order in which the tasks and subtasks are executed, and multiple preset responses associated with the subtasks. A response generation unit analyzes the meaning of the utterance text by referring to a conversation history consisting of utterance text and response text based on the user's utterances, selects an appropriate preset response in the utterance text, and generates a response text based on the selected preset response. It includes a response audio signal generation unit that synthesizes speech based on the response text and generates a response audio signal, The aforementioned response generation unit, The system searches for a preset answer associated with the currently running subtask based on the spoken text. If there is no suitable preset answer for the currently running subtask, it searches for a preset answer within the same task. If there is no suitable preset answer within the same task, the system will search for preset answers in other tasks. A response generation device characterized by the following features.

2. The aforementioned response generation unit, The system remembers whether the selection of preset answers for subtasks and tasks is complete or incomplete. If there is no suitable preset answer for the currently running subtask, the system will first search for incomplete preset answers within the same task, and then search for completed preset answers within the same task. If there is no suitable preset answer within the same task, search for preset answers in other incomplete tasks, and then search for preset answers in other completed tasks. The answer generation device according to feature 1.

3. The aforementioned response generation unit, Search for preset answers within all tasks that need to be searched, and if a suitable answer cannot be selected, perform error handling. The answer generation device according to feature 1.

4. The aforementioned response generation unit, As part of the error handling, a request to switch to another device is executed. The answer generation device according to feature 3.

5. The answer generation device according to claim 1, The aforementioned utterance text and the aforementioned response text are displayed on the display unit and monitored by a monitoring operator. The conversation monitoring device includes, when any of the following are detected as trouble factors: repetition of negative words in the utterance text, repetition of answers with the same meaning in the response text, or determination that an appropriate answer cannot be given, the display unit displays a blank space where text can be entered and an alert indicating that a trouble factor has been detected, and supplies the input text entered by the monitoring operator as the response text to the response generation device. A response generation system characterized by the following features.

6. It has a storage unit that stores the utterance text and the response text, and the input text entered by the monitoring operator, Add preset answers based on the input text. The answer generation system according to claim 5, characterized by having the following features.

7. For computers, A response generation program characterized by executing [the specified action].

8. A method for generating answers, characterized by having the following features.

9. A CTI device that controls the line connection between the user terminal and the call center operator, and records the conversation history between the user terminal and the call center operator, A customer management device that manages customer information, including customer personal information and attributes, as well as usage history, by associating it with a customer ID (IDentification), A status management device that manages the current usage status of all customers, A speech-to-text generation unit that converts the user's voice into text and generates speech text, A storage unit that stores multiple tasks to be performed as specified in a conversation with the user, multiple subtasks associated with those tasks and which are details for performing those tasks, the order in which the tasks and subtasks are executed, and multiple preset responses associated with the subtasks. A response generation unit analyzes the meaning of the utterance text by referring to the utterance text and conversation history, selects an appropriate preset response in the utterance text, and generates a response text based on the selected preset response. It includes a response audio signal generation unit that synthesizes speech based on the response text and generates a response audio signal, The aforementioned response generation unit, The system searches for a preset answer associated with the currently running subtask based on the spoken text. If there is no suitable preset answer for the currently running subtask, it searches for a preset answer within the same task. If there is no suitable preset answer within the same task, the system will search for preset answers in other tasks. An AI call center system characterized by the following features.

Images